Introduction to Non-Standard Analysis

Home , Abraham Robinson, Internal set, Nonstandard analysis, Transfer principle

INTRODUCTION TO NON-STANDARD ANALYSIS

DIEGO ANDRES BEJARANO RAYO

Abstract. In this paper we introduce the basic concepts of non-standard analysis. First, we provide an overview of first order logic and the theory of filters to successfully construct the set of hyperreal numbers, which we will use as our object of study. Using the transfer principle, a simple corollary ofLo´s’sTheorem, we not only introduce the non-standard notions of continuity, differentiability and Riemann integrability for functions of one variable, but also show that they are equivalent to the standard concepts taught in a first-year calculus course.

Contents 1. First Order Logic. 1 2. Filters, Ultraﬁlters, Ultraproducts and Ultrapowers 3 3. The Structure of the Ultraproduct 6 4. The Structure of the Hyperreals 7 5. Continuity and Diﬀerentiability 11 6. Integration 14 7. Conclusion 17 Acknowledgments 17 References 17

1. First Order Logic. Before making an introductory exposition of the ideas behind non-standard analysis and the related construction, it is important that the reader familiarizes himself with the theory of first order logic. This section presents an introduction to these ideas, with a few examples. The system of first order logic is defined inductively, starting by defining the alphabet or the collection of symbols that will be used.1 Definition 1.1. The alphabet of first order logic is a set containing the following elements:

• An infinite list of constant symbols: a, b, c, a1, b1, c1,... • An infinite list of variable symbols: x, y, z, x1, y1, z1,... • An infinite list of function symbols: f, g, h, f1, g1, h1,... • An infinite list of relation symbols: P, R, Q, P1,R1,Q1,... • Logical connectors: ∧, ∨, ¬, →, ↔, with their usual interpretation. • Logical quantifiers: ∀, ∃. • The equality symbol: =. • The two parentheses “(” and “)”. Note that every function and relation symbol is an n-placed function or relation symbol. This number is commonly referred to as the arity of the function or relation. For example, a function with arity 1 is called unary, with arity 2 is called binary, so on and so forth. We will denote the arity of a function as ar(f), so that for a binary function f, for example, we have ar(f) = 2.

Deﬁnition 1.2. A term is a string of symbols from the alphabet that is deﬁned recursively as follows:

1The material presented in this section is based off of [1] and [2]. 1 (1) Every constant is a term. (2) Every variable is a term. (3) If f is a function with arity n and t1, . . . , tn are terms then f(t1, . . . , tn) is also a term. (4) A string of symbols is a term if it can be constructed applying the previous steps finitely many times. At this point, it becomes important to point out that only the ∧, ¬ and ∃ logical symbols are worth mentioning, since all the other logical connectors can be derived from these three. Note that (ψ ∨ ϕ) = ¬((¬ψ) ∧ (¬ϕ)), (ψ → ϕ) = ¬(ψ ∧ (¬ϕ)), (ψ ↔ ϕ) = ¬(ψ ∧ (¬ϕ)) ∧ ¬(ϕ ∧ (¬ψ)) and (∀x)ϕ = ¬(∃x¬ϕ). Therefore, from now on, we will only mention the three necessary connectors, and assume that the rest follow from these equalities. Definition 1.3. A formula is a string of symbols from the alphabet that is defined recursively as follows:

(1) If t1 and t2 are terms then (t1 = t2) is a formula. (2) If R is a relation with arity n and t1, . . . , tn are terms then (R(t1, . . . , tn)) is a formula. (3) If ϕ is a formula then so is ¬ϕ. (4) If ϕ and ψ are formulas, then so is (ϕ ∧ ψ). (5) If ϕ is a formula and x is a variable then (∃x)ϕ is also a formula. (6) A string of symbols is a formula if it can be constructed by finitely many applications of the previous steps. This also implies that, for example, (ψ ∨ ϕ) is a formula if, and only if, ψ and ϕ are formulas by using the fact that (ψ ∨ ϕ) = ¬((¬ψ) ∧ (¬ϕ)) and the definition above. Remark 1.4. The formulas described in 1.3.1 and 1.3.2 are called atomic formulas. Also, note that (1) can be understood as a binary relation R=(t1, t2) which is true if, and only if, t1 = t2. One of the limitations of first order logic is that it does not allow quantification over relations, only variables. This is the key difference between first and higher order logic. While the definitions in this section will focus on developing first order logic, this distinction between first order and higher logic statements will be key in the development of non-standard analysis.

Next, we will discuss the difference between free and bound variables. In not so rigorous terms, we say that a variable is free if it doesn’t appear next to a quantifier and bound otherwise. Definition 1.5. Let ϕ be a formula. We define the set of free variables of ϕ, denoted as FV (ϕ), inductively as follow:

(1) If ϕ = (t1 = t2), then FV (ψ) = {x | x appears in t1 or t2} (2) If ϕ = (R(t1, . . . , tn)) for some relation of arity n, then FV (ϕ) = {x | x appears in ti for some 1 ≤ i ≤ n}. (3) If ϕ = (¬ψ), where ψ is a formula, then FV (ϕ) = FV (ψ). (4) If ϕ = (µ ∧ ν), where µ and ν are formulas, then FV (ϕ) = FV (µ) ∪ FV (ν). (5) If ϕ = (∃x)ψ, where ψ is a formula, then FV (ϕ) = FV (ψ) \{x}. Definition 1.6. A formula ϕ is called a sentence if it has no free variables, meaning FV (ϕ) = Ø. Definition 1.7. A language L is a set containing all logical symbols and quantifiers (including the equality sign and the parenthesis) and some arbitrary number of constants, variables, function symbols and relation symbols. It is understood that all formulas made from any language L follow the previous rules. Definition 1.8. Let A be some nonempty set and V ⊂ L be the set of all variables in L.A variable assignment is a mapping β : V → A, which assigns elements of A to all variables in V . Particularly, for some element k ∈ A, some variable x ∈ V and some variable assignment function β, there is a function β[x, v] defined as ( k if x = y β[x, k](y) = β(y) if x 6= y

2 Definition 1.9. A model or structure M for some language L is an ordered triple M = (A, I, β), where A is a nonempty set, β is a variable assignment function and I is an interpretation function with domain the set of all constants, relations and function symbols in L such that: (1) For every constant symbol c ∈ L, we have that I(c) ∈ A. (2) For every function symbol f ∈ L with arity n, we have that I(f) ∈ An × A. Meaning I(f) is a function of arity n defined on A. (3) For every relation symbol R ∈ L with arity n, we have that I(R) ⊂ An. Meaning I(R) is the set of all n-tuples that satisfy R under I. A is frequently called the universe of M. Note that most authors define models as the pair (A, I) and leave the variable assignment function out, even when some have defined such a function. The reason why we decide to define models as a triple is simple; it makes the following definition easy to understand and does not cause any real change on what we consider a model to be. Definition 1.10. Let L be a language and M = (A, I, β) a model of L. Then the interpretation of any term t, denoted as (t)I,β, of symbols in L is defined as follows: • If t = c for some constant c, then (t)I,β = I(c). • If t = x for some variable x, then (t)I,β = β(x). I,β I,β I,β • if t = f(t1, . . . , tn) for some function f of arity n, then (t) = I(f)((t1) ,..., (tn) ). Definition 1.11. Let L be a language, M = (A, I, β) a model for L and ϕ some formula in L. Then, we say that M satisfies ϕ and write M |= ϕ or (A, I, β) |= ϕ whenever:

• If ϕ = R(t1, . . . , tn) for some relation R ∈ L of arity n, meaning ϕ is atomic, then (A, I, β) |= ϕ if I,β I,β ((t1) ,..., (tn) ) ∈ I(R). • If ϕ = ¬ψ for some atomic formula ψ, then (A, I, β) |= ϕ if (A, I, β) does not satisfy ψ. • If ϕ = (µ∧ν) for some atomic formulas µ and ν, then (A, I, β) |= ϕ if (A, I, β) |= µ and (A, I, β) |= ν. • If ϕ = (∃x)ψ for some atomic formula ψ, then (A, I, β) |= ϕ if there exists some k ∈ A such that (A, I, β[x, k]) |= ψ. Note that in this case we assume that x is a free variable on ψ. This concludes the exposition of ﬁrst order logic. While the reader might feel like the next section is quite disconnected with the previous one, both are essential in constructing the basic framework of non-standard analysis.

2. Filters, Ultrafilters, Ultraproducts and Ultrapowers Filters are a way of formalizing the notion of “big” within set theory. The following definitions set the groundwork for the study of filters, ultrafilters, ultraproducts, ultrapowers and their application within logic.2 Definition 2.1. A filter F on a set I is a set F ⊂ P(I) such that i. I ∈ F ii. If X ∈ F and X ⊂ Y , then Y ∈ F for all X,Y ∈ P(I). iii. If X,Y ∈ F , then X ∩ Y ∈ F for all X,Y ∈ P(I). Intuitively, we are stating that the sets in F are considered to be “big,” while those not in it are not. While this notion can be quite arbitrary, as we’ll see in the next few examples, we are setting a few rules to this notion of “big”: the entire set must be big, or else nothing in that set should be; sets which are bigger than “big” sets are also “big” and finite intersections of “big” sets are “big.” While the final condition might not be quite as intuitive as the first two, it proves to be quite useful and can be explained by stating that sets are actually “big” if their intersections are also “big.” Remark 2.2. {I} and P(I) are filters of I for any set I. The proof of this remark should be straightforward and is thus left to the reader. Note that the filters in 2.2 are called the trivial filters of I. A non-trivial filter is called proper.

Deﬁnition 2.3. Let x ∈ A, then Fx = {Y ∈ P(I) | x ∈ Y } is called the principal ﬁlter of x over I.

2The material presented in Sections 2 and 3 is based off of [3], [4], [5] and the Logic lectures by Denis Hirschfeldt and Maryanthe Malliaris during the 2015 Math REU at the University of Chicago. 3 Lemma 2.4. For any set x ∈ A, the principal filter of x over I is a filter of I.

Proof. The proof of this lemma is quite simple. Since x ∈ I, then I ∈ Fx. If x ∈ X ∈ Fx and X ⊂ Y , it must follows that x ∈ Y , so Y ∈ Fx. Finally Fx is closed under finite intersection since x ∈ X and x ∈ Y implies X ∈ X ∩ Y , so Fx is a filter. Definition 2.5. A non-trivial and non-principal filter of I is called a free filter. Theorem 2.6. A filter F over some set I is free if, and only if, T A = Ø A∈F T Proof. ⇒ Take the contrapositive. Assume that A∈F A 6= Ø. Therefore, there exists some set x ∈ I such that x ∈ Y for all Y ∈ F . Thus, F ⊂ {Y ∈ P(I) | x ∈ Y }. Furthermore, since F is a filter, it is upwardly closed, so it must contain every set containing x and, hence, F = {Y ∈ P(I) | x ∈ Y }. So F is principal. ⇐ Let T A = Ø and assume towards contradiction that F is principal. Then there is some x ∈ I such A∈F T T that F = {Y ∈ P(I) | x ∈ Y }. This implies that A∈F A = {x}, but we assume that A∈F A = Ø. This gives us that {x} = Ø. Contradiction. Remark 2.7. Since filters are closed under finite intersection, 2.6 implies there are no free filters on finite sets.

Definition 2.8. The FréchetFilter on some set I is defined as FI = {X ⊂ I | X is cofinite}. Lemma 2.9. Let I be an infinite set. Then the FréchetFilter over I is a free filter. Proof. Assume towards contradiction that the Fréchet filter over I is principal. Then, by 2.6 we have that T T A∈F A 6= Ø. Fix some k ∈ A∈F A and some X ∈ F. This implies that k ∈ X. However, note that k 6∈ X \{k}. Since X is cofinite, X \{k} is also cofinite, which gives us that X \{k} ∈ F. Therefore, T k 6∈ A∈F A. Contradiction. Lemma 2.10. Every free filter contains the Fréchetfilter.

Proof. Let F be a free filter over some infinite set I. Let FI denote the Frechet filter over I. Fix some Y ∈ FI , then it follows that I \Y is a finite set. Since F is a free filter, for every x ∈ I \Y there exists some set T Kx ∈ F such that x 6∈ Kx. Since F is closed under finite intersection, it follows that K = x∈I\Y Kx ∈ F and K ⊂ Y . This implies that Y ∈ F since F is upwardly closed. Thus, Fi ⊂ F . Definition 2.11. A filter U on a set I is an ultrafilter of I if for all X ⊂ I either X ∈ U or I \ X ∈ U but not both. Theorem 2.12. Let x ∈ A, then the principal filter of x over I is an ultrafilter of I.

Proof. Let Fx be the principal filter of x over I. By previous observations, Fx is a filter. Next, let Y be an arbitrary subset of I. Then x ∈ Y or x 6∈ Y ; if x ∈ Y then x 6∈ I \ Y so Y ∈ Fx and I \ Y 6∈ Fx. Conversely, if x 6∈ Y , then x ∈ I \ Y so Y 6∈ Fx and I \ Y ∈ Fx. Either way, only one of the sets is in the filter. Thus, Fx is an ultrafilter by 2.11. Note that it is consistent with ZF set theory that there are not non-principal ultrafilters. In this paper, we require the Axiom of Choice to prove the existence of non-principal ultrafilters. Definition 2.13. A set G ⊂ P(I) has the finite intersection property (FIP) if the intersection of any finite number of elements of G is nonempty. Remark 2.14. Note that every filter has the finite intersection property. Theorem 2.15. Every S ⊂ P(I) with FIP has a proper filter containing it. We call this filter the filter generated by S. Proof. Let F = T {F ⊂ P(I) | S ⊂ F and F is a proper filter on I}. Note that Ø 6∈ F and I ∈ F since F is the intersection of proper filters and S has the FIP. Let X ∈ F and X ⊂ Y . It follows that X is an element of every filter and, therefore, so is Y ; this implies that Y ∈ F . Finally, let X,Y ∈ F ; by similar argument it follows that X ∩ Y ∈ F . Thus, F is a filter on I and S ∈ F by definition. 4 Lemma 2.16. (Ultrafilter) Every proper filter E ⊂ P(I) is contained in some ultrafilter U ⊂ P(I).

Proof. Let F = {F ⊂ P(I) | F ⊃ E is a proper filter of I}. Note then that F is a partially ordered set with ⊂ as a partial order relationship. Let C ⊂ F be an arbitrary chain (totally ordered subset) of F ; we want to show that C is bounded. Consider S C, it is clear that every element in C is a subset of S C, so S C is a bound of C. Thus, it suffices to show that S C ∈ F to show that F has a maximum element by Zorn’s lemma. Since every element in C is a subset of P(I), it follows that S C ⊂ P (I). Note also that I ∈ S C. Next consider some X ∈ S C and let X ⊂ Y ⊂ P(I); since X ∈ S C there exists some filter F such that X ∈ F . Therefore, Y ∈ F as F is upwardly closed and, hence, Y ∈ S C. Finally, let X,Y ∈ S C, then there are filters F,F 0 such that X ∈ F and Y ∈ F 0; since C is a chain we can assume without loss of generality that F ⊂ F 0, which implies that X,Y ∈ F 0. This gives us that X ∩ Y ∈ F 0 and, hence, X ∩ Y ∈ S C. Hence S C is a filter, so S C ∈ F . Thus, by Zorn’s lemma, F contains a maximal element, denote this element by U. Finally, we want to show that U is an ultrafilter; it is clear that Ø 6∈ U so U cannot contain both a subset of P(I) and its complement. Assume for contradiction that U is not an ultrafilter; then there is some set A ⊂ I such that A 6∈ U and I \ A 6∈ U. Consider then U ∪ {A}, it follows that U ∪ {A} has a finite intersection property since U has a finite intersection property and I \ A 6∈ U (so no subset of I \ A is an element of U). Let S be the filter generated by U ∪ {A}. It is clear that U ⊂ S, which implies that U is not maximal in F . Contradiction.

Corollary 2.17. There exists some free ultrafilter on N. Proof. Let F be the Frechet filter on N. Since N is infinite such a filter exists. By 2.16 there exists some N T T T ultrafilter of N, denoted by U, containing F . Thus, A∈U A ⊂ A∈F A. This implies that A∈U A = Ø N N and, hence, U is a free ultrafilter.

From this point on we assume I is an inﬁnite set with some free ultraﬁlter U and {Ai}i∈I is a collection of nonempty sets.

Definition 2.18. Let U be a free ultrafilter on some indexing set I and let {Ai}i∈I be a collection of nonempty sets. Then the arbitrary product of the collection is defined as Y Ai := {f | f has domain I and f(i) ∈ Ai for all i ∈ I} i∈I Q Definition 2.19. Two functions f, g ∈ i∈I Ai are modulo U equivalent if {i ∈ I | f(i) = g(i)} ∈ U. We write f =U g to indicate this relationship. Lemma 2.20. Modulo U equivalence is an equivalence relation.

Proof. It suﬃces to show that =U is reﬂexive, symmetric and transitive. Q • Consider some function f ∈ i∈I Ai. Note that {i ∈ I | f(i) = f(i)} = I ∈ U,

which implies that f =U f. Q • Let f, g ∈ i∈I Ai. Then if f =U g, we have {i ∈ I | f(i) = g(i)} ∈ U, but notice that {i ∈ I | f(i) = g(i)} = {i ∈ I | g(i) = f(i)} ∈ U,

so g =U f, which implies =U is reflexive. Q • Finally, let f, g, h ∈ i∈I Ai be such that f =U g and g =U h. Then K = {i ∈ I | f(i) = g(i)} ∈ U and L = {i ∈ I | g(i) = h(i)} ∈ U. Since U is closed under finite intersection, we have that K ∩ L = {i ∈ I | f(i) = g(i) = h(i)} ∈ U, which gives us that f =U h. Thus, =U is an equivalence relation. Q Definition 2.21. We define [f]U as the equivalence class of all functions g ∈ Ai such that g =U f for Q i∈I some function f ∈ i∈I Ai. 5 Definition 2.22. The ultraproduct of {Ai}i∈I modulo U is Q Y Ai /U := {[f]U | f ∈ Ai} i∈I i∈I

Definition 2.23. If we let Ai = A for some set, then ultrapower of A modulo U is Q Y A /U = {[f]U | f ∈ A} i∈I i∈I Definition 2.24. Let U be a free filter of N, then the set of hyperreal numbers, denoted as ∗R, is defined as the ultrapower of R modulo U. In other words: ∗ Q R = R /U n∈N This definition implies that the elements of the Hyperreals are the equivalence classes of sequences under modulo U equivalence.

3. The Structure of the Ultraproduct One of the immediate questions that arises from the ultraproduct construction is “How is the Ultraproduct of a collection of sets similar, or different, to the sets in that collection?” This section focuses on introducing the structure of the Ultraproduct, using the tools introduced in Section 1, and provides a proof forLo´s’s Theorem, which will be key to the development of non-standard analysis. Definition 3.1. Let L be a language, then a theory of L is a set of sentences of L. Definition 3.2. Let L be a language, M a model for L and T a theory. We say that M satisfies T , and write M |= T if M |= ϕ for all ϕ ∈ T . Definition 3.3. Let L be a language, M a model for L. The theory of M, denoted as T h(M), is the set of all sentences ϕ of L such that M |= ϕ.

Definition 3.4. Let I be an index set with some ultrafilter U on I and let Mi = (Ai,Ii, βi) be a model for ∗ Q ∗ ∗ some language L for all i ∈ I. Then the ultraproduct M = i∈I Ai /U, I, β is a model of L with an interpretation function ∗I and variable assignment function ∗β defined as follows: ∗ • If x is a variable in L, then β(x) = [(βi(x))]U. ∗ • If c is a constant in L, then I(c) = [(Ii(c))]U. ∗ • If f is a function symbol of arity n, then I(f)([g1]U,..., [gn]U) = [(Ii(f)(g1(i), . . . , gn(i)))]U ∗ • If R is a relation symbol of arity n, then ([g1]U,..., [gn]U) ∈ ( I)(R) if, and only if, {i ∈ I | (g1(i), . . . , gn(i)) ∈ Ii(R)} ∈ U. It is worth noting that in the definition above we make use of sequence notation. Remark 3.5. The previous definition is well defined, meaning that it does not depend on the choices of the [gi]U. Q 0 0 ∗ Proof. Fix g1, . . . gn, g1, . . . gn i∈I Ai such that g1 =U g1, . . . , gn =U gn and ([g1]U,..., [gn]U) ∈ ( I)(R) for 0 0 some relation symbol R of arity n. Notice then that S = {i ∈ I | g1(i) = g1(i) ∧ ... ∧ gn(i) = gn(i)} ∈ U since U is closed under finite intersections. This implies that 0 0 (∀i ∈ S)((g1(i), . . . , gn(i)) ∈ Ii(R) ⇔ (g1(i), . . . , gn(i)) ∈ Ii(R)) 0 0 which gives us that {i ∈ I | (g1(i), . . . , gn(i)) ∈ Ii(R)} ∈ U since it is a superset of S. This implies that 0 0 ∗ ([g1]U,..., [gn]U) ∈ ( I)(R). So R is well-defined.

0 0 In the case of functions, we get a similar result. Consider again some functions g1, . . . gn, g1, . . . gn ∈ Q i∈I Ai as above and some function symbol f of arity n. Deﬁne the set S as above and notice that S ∈ U, then 0 0 (∀i ∈ S)(Ii(f)(g1(i), . . . , gn(i)) = Ii(f)(g1(i), . . . , gn(i)) 6 since f is a function and all the inputs are the same. Therefore, the sequence (Ii(f)(g1(i), . . . , gn(i))) is 0 0 modulo U equivalent to the sequence (Ii(f)(g1(i), . . . , gn(i)) so both belong to the same equivalence class. This implies that f is well deﬁned.

Theorem 3.6. (Lo´s’sTheorem) Let L be a language, I be a set with some ultrafilter U on I and Mi = ∗ Q A ∗ ∗ (Ai,Ii, βi) be a model for L for all i ∈ I. Then for all ϕ of L we have that M = i∈I i /U, I, β |= ϕ if, and only if, {i ∈ I | Mi |= ϕ} ∈ U. Proof. We proceed inductively on the complexity of ϕ: (1) If ϕ is an atomic formulae, then the statement holds by the previous definition. (2) If ϕ = (µ ∧ ν), where µ and ν are atomic formulas, then: ⇒ If ∗M |= µ ∧ ν, then ∗M |= µ and ∗M |= ν by definition. Therefore, since µ and ν are atomic, it follows that {i ∈ I | Mi |= µ} ∈ U and {i ∈ I | Mi |= ν} ∈ U. Since U is closed under finite intersection, we have that {i ∈ I | Mi |= µ ∧ Mi |= ν} = {i ∈ I | Mi |= (µ ∧ ν)} ∈ U. ⇐ If {i ∈ I | Mi |= (µ ∧ ν)} ∈ U, note that {i ∈ I | Mi |= (µ ∧ ν)} ⊂ {i ∈ I | Mi |= µ} and {i ∈ I | Mi |= (µ ∧ ν)} ⊂ {i ∈ I | Mi |= ν}. Since U is upwardly closed, it follows that both sets are in U, so ∗M |= µ and ∗M |= ν, which gives us that ∗M |= µ ∧ ν. (3) If ϕ = (¬ψ), where ψ is an atomic formula, then if ∗M |= ϕ, we have that ∗M does not model ψ, which implies that {i ∈ I | Mi |= ψ} 6∈ U since ψ is atomic. This gives us that {i ∈ I | Mi does not model ψ} ∈ U since U is an ultrafilter. However, note that {i ∈ I | Mi does not model ψ} = {i ∈ I | Mi |= ϕ} ∈ U. Note that all the steps in this part of the proof are reversible, so this proves the biconditional. (4) If ϕ = (∃x)ψ, where ψ is an atomic formula and x a free variable in ψ, then: ∗ Q Q ∗ ∗ ⇒ If M |= (∃x)ψ, we have that there is a [g]U ∈ i∈I Ai /U such that ( i∈I Ai /U, I, β[x, [g]U]) |= ψ. Since ψ is atomic, this implies that {i ∈ I | (Ai,Ii, βi[x, g(i)]) |= ψ} ∈ U. By definition, it follows that {i ∈ I | (Ai,Ii, βi[x, g(i)] |= ψ} = {i ∈ I | Mi |= (∃x)ψ} ∈ U. S ⇐ If {i ∈ I | Mi |= (∃x)ψ} ∈ U, define a function g : I → i∈I Ai such that for all i ∈ {i ∈ I | Mi |= (∃x)ψ} we have that g(i) is such that (Ai,Ii, βi[x, g(i)]) |= ψ and g(i) ∈ Ai otherwise. Note that such g(i) exist by assumption but this step also requires the axiom of choice. Furthermore, it is clear that Q Q ∗ ∗ [g]U ∈ i∈I Ai /U by the definition of the ultraproduct. Therefore, ( i∈I Ai /U, I, β[x, [g]U]) |= ψ by the definition of g and this in turn gives us that ∗M |= (∃x)ψ. Since all formulas of L are obtained by finite application of the steps above, the proof is complete An important corollary ofLo´s’sTheorem is known as the Transfer Principle, and it has to do with how the first order Theory of Ultrapowers relates to the first order theory of the set that makes up the ultrapower. Corollary 3.7. (Transfer Principle) Let L be a language, I be a set with some ultrafilter U on I and ∗ Q ∗ ∗ M = (A, I, β) be a model for L. Then for all ϕ of L we have that M = i∈I A /U, I, β |= ϕ if, and only if, M = (A, I, β) |= ϕ. In other words, ∗M |= T h(M).

∗ Proof. This statement follow directly fromLo´s’stheorem. If M |= ϕ, then {i ∈ I | Mi |= ϕ} ∈ U, but since all the Ai = A, it follows that {i ∈ I | Mi |= ϕ} = I since otherwise it would equal the empty set. This ∗ implies that M |= ϕ. Conversely, if M |= ϕ, then {i ∈ I | Mi |= ϕ} = I ∈ U, so M |= ϕ byLo´s’stheorem.

4. The Structure of the Hyperreals It follows from 3.7 that ∗R satisfies the same first order theory that R satisfies. This implies first that if we define addition and multiplication as determined by 3.4, then ∗R is a field. While this last statement can be proven withoutLo´s’stheorem (or in this case 3.7), we will approach it using this particular result. Note that it suffices to show that all field axioms can be expressed as first order logic statements, meaning using only the rules introduced in Section 1 of this paper, since 3.4 and 3.7 prove that ∗R must also satisfy these statements.3

3From this point on the paper relies heavily on the works of [5] and [6]. 7 Theorem 4.1. ∗R is a field. ∗ Proof. For the remainder of this proof let x, y, z ∈ R. This implies that x = [(xn)]U, y = [(yn)]U and z = [(zn)]U for some sequences (xn), (yn), (zn). Furthermore, following 3.4 we define addition and multiplication point-wise, so that x + y = [(xn + yn)]U and xy = [(xnyn)]U. These operations are well-defined as shown in 3.5. Also, note that addition and multiplication are closed binary operations on ∗R since the sum or product of two sequences is also a sequence. Thus, it suffices to show that the field axioms can be written as first order statements: (1) Commutativity of Addition: (∀x, y)(x + y = y + x) (2) Associativity of Addition: (∀x, y, z)(x + (y + z) = (x + y) + z) (3) Existence of an Additive Identity: (∃0)(∀x)(x + 0 = 0 = 0 + x) (4) Existence of an Additive Inverse: (∀x)(∃y)(x + y = 0 = y + x) (5) Commutativity of Multiplication: (∀x, y)(x · y = y · x) (6) Associativity of Multiplication: (∀x, y, z)(x · (y · z) = (x · y) · z) (7) Existence of a Multiplicative Identity: (∃1)(∀x)(x · 1 = 1 = 1 · x) (8) Existence of a Multiplicative Inverse: (∀x)(∃y)(¬(x = 0) → (x · y = 1 = y · x)) (9) Distributivity of Multiplication over Addition: (∀x, y, z)(x · (y + z) = (x · y) + (x · z)) (10) Distinctness of Identities: ¬(0 = 1) While the reader might already be familiar with these statements and with the fact that these are indeed first order logic statements, this should provide sufficient proof of this claim. Note that we did not use the notation +(a, b) and instead chose to write (a + b); this is done mostly due to convention and does not affect ∗ the validity of the statements. Thus, by 3.7, it follows that R satisfies all the above and it is a field. ∗ Furthermore, we can also show that R is totally ordered by a relation ≤U. To define this relation, we again appeal to 3.4 and say that x ≤ y if, and only if, {n ∈ N | xn ≤ yn} ∈ U. It should be clear that the fact that ∗R is totally ordered follows from the fact that R is totally ordered and that U is an ultrafilter. However, we can also use 3.7 by showing that the order axioms can be written using first order logic. Thus, we have that ∗R is a totally ordered field. However, we will show that ∗R is not complete, as it does not satisfy the least-upper-bound property. To do so, we will first introduce some important relationships between R and ∗R. Theorem 4.2. There exists an embedding of R into ∗R. That is, there exists a function i: R → ∗R that is injective and preserves addition, multiplication, and ordering.

∗ Proof. Define i: R → R as i(r) = [(r, r, r, ...)]U. To show that i is injective, consider the contrapositive of injectivity. If r 6= r0, then the sequences (r, r, r, ...) and (r0, r0, r0, ....) are not modulo U equivalent as they differ in all entries. To show that i respects addition and multiplication, meaning that i(r) + i(s) = i(r + s) and that i(r) · i(s) = i(r · s) (where addition and multiplication are done in the respective sets), it suffices to see that

i(r) + i(s) = [(r, r, r, ...)]U + [(s, s, s, ...)]U = [(r + s, r + s, r + s, ....)]U = i(r + s) and that the corresponding statement for multiplication is the same. To show that i respects ordering, meaning that if r ≤ s then i(r) ≤ i(s), let r ≤ s and then note that [(r, r, r, ...)]U ≤ [(s, s, s, ...)]U since {n ∈ N | r ≤ s} = N ∈ U. This embedding will be crucial in our development of nonstandard analysis since it will allow us to represent real numbers as hyperreals. Also, note that i also follows directly from 3.4, which is also a crucial insight. Next, we will prove the existence of two kinds of hyperreals that will make nonstandard analysis possible: unlimited and inﬁnitesimal numbers.

Theorem 4.3. There exists a hyperreal number ω such that |ω| > i(r) for all r ∈ R+. Similarly, there exists a hyperreal number ε such that 0 < |ε| < i(r) for all r ∈ R+. Note that we define absolute value in ∗R using 3.4. Furthermore, the absolute value function holds all of the same properties it has in R, including the triangle inequality. The proof of this comes from applying this property to every entry in the sequence and then using the properties of U as an ultrafilter. 8 Proof. Let ω = [(an = n)]U, so ω = [(1, 2, 3, ...)]U. Fix some r ∈ R, note that since R is Archimedean, we have that r ≤ m for some m ∈ N. Let i(m) = [(mn = m)]U; it follows that for all n ≤ m we have an ≤ m. However, this implies that {n ∈ N | an ≤ mn} is finite, so {n ∈ N | an > mn} ∈ U since U contains all the cofinite sets. Thus, ω > i(m) ≥ i(r). Since r was chosen arbitrarily, his holds for all r ∈ R+. Next, let 1 1 −1 ∗ −1 −1 ε = [(1, 2 , 3 , ...)]U = ω . Since R is a field it follows that ε = ω < i(r) , but since this holds for all + + r ∈ R , we also get that ε < i(r) for all r ∈ R . This motivates the following classification for hyperreal numbers:

Definition 4.4. A hyperreal number ω is said to be unlimited if |ω| > i(r) for all r ∈ R+. Definition 4.5. A hyperreal number ε is said to be infinitesimal if |ε| < i(r) for all r ∈ R+. Definition 4.6. A hyperreal number n is said to be finite if there exists r, s ∈ R+ such that i(r) ≤ |n| ≤ i(s). Definition 4.7. A hyperreal number x is said to be limited if it is not unlimited. For example, the ω presented in 4.3 is unlimited and the ε presented in 4.3 is infinitesimal. Also of note is that 0 ∈ R is the only real number whose image under i is infinitesimal.

It is important to consider then how addition and multiplication of hyperreals plays with these definitions. As a general rule, the addition of any hyperreal to an unlimited hyperreal is also unlimited, the addition of any limited hyperreal to a finite hyperreal is also finite, and only the addition of two infinitesimals is infinitesimal. As for the products, the product of a finite hyperreal with and infinitesimal is infinitesimal; the product of two finite hyperreals is finite, and the product of an unlimited hyperreal and an unlimited or a finite hyperreal is also unlimited. Finally, the product of an unlimited hyperreal and an infinitesimal follows no general rules and depends on the numbers themselves. The proofs for all of these claims follow from the definitions given above. With this in mind, we can show that ∗R is not complete: Theorem 4.8. i(R) ⊂ ∗R is a nonempty bounded set without supremum. Proof. Note that i(R) is bounded above by any unlimited hyperreal, and it is of course nonempty. However, assume for contradiction that i(R) has a supremum and set ω = sup i(R). Then ω must be unlimited, as otherwise it wouldn’t be an upper bound, but this also implies that ω − i(1) is unlimited and, hence, an upper bound of i(R). However, it is clear that ω − i(1)

Deﬁnition 4.11. For any hyperreal number x0, we deﬁne the halo, sometimes called the monad, of x0 as ∗ the set of all hyperreals that are close to x0. More formally, hal(x0) = {x ∈ R | x ' x0}.

Definition 4.12. For any hyperreal number x0, we define the galaxy of x0 as the set of all hyperreals x ∗ such that x − x0 is limited. More formally, gal(x0) = {x ∈ R | x − x0 is limited}. 9 A few important remarks come from this last definition. Note that the set of all infinitesimals is precisely hal(0) and, thus, we will normally denote a number as infinitesimal by writing x ' 0. It is clear that for 0 we mean i(0), but to ease notation we will omit the i whenever it is clear that the number should be hyperreal. Also, we will denote unlimited numbers using the notation ω ' ∞, which is meant to capture the notion that unlimited numbers are greater than all real numbers. It is worth noting that we have a way to relate real numbers to specific hyperreal numbers (via the embedding defined above), but so far we have no way of relating arbitrary hyperreal numbers to some specific real number. The next theorem is meant to resolve that issue: Theorem 4.13. Given any limited hyperreal x, there exists a unique real number r, called the shadow or standard part of x, such that i(r) ' x. We denote this relation by writing sh(x) = r.

Proof. Fix some hyperreal x and let A = {r ∈ R | i(r) < x}. Since x is limited, it follows that A is nonempty and bounded above; thus, A attains supremum. Let a = sup A; we claim that a = sh(x). To show this, ﬁx ε ∈ R+ and consider k − i(a); since a = sup A, it follows that 0 ≤ k − i(a), which implies that i(ε) < k − i(a). Notice that since a = sup A and ε > 0, it follows that k < i(a + ε), which gives us that k − i(a) < i(ε). Thus, |k − a| < i(ε). Since this holds for all ε ∈ R+, we have that k ' i(a). To prove that a is indeed unique, assume that k ' i(a) and k ' i(a0). This implies that i(a) ' i(a0). Thus, 0 0 0 i(a − a ) is inﬁnitesimal, but note that this is only possible if a − a = 0. Thus, a = a . Theorem 4.14. The shadow operator respects addition, multiplication and ordering.

Proof. First we will show that the shadow operator respects addition. Fix any a, b ∈ ∗R, we want to show that sh(a) + sh(b) = sh(a + b). Let sh(a) = a and sh(b) = d; thus, a ' i(c) and b ' i(d). It suffices to + i(ε) i(ε) show that a + b ' i(c) + i(d). Fix ε ∈ R and notice that |a − i(c)| < 2 and |b − i(d)| < 2 , by triangle inequality we get that |(a + b) − (i(c) + i(d)| < i(ε), which shows that a + b ' i(c + d) and thus completes the proof. To show that sh(a)sh(b) = sh(ab) we again let a ' i(c) and b ' i(d) and fix ε ∈ R+ and let i(ε) i(ε) |a − i(c)| < min 1, and |b − i(d)| < 2(|i(d)| + 1) 2(|i(c)| + 1) This implies that |ab − i(c)i(d)| < i(ε), so ab ' i(c)i(d). The proof of this statement follows similarly from the proof of this same statement in R, which should not be foreign to the reader, and applying the properties of U as an ultrafilter. Finally, we want to show that if a < b, then sh(a) ≤ sh(b). We have two cases, either a ' b or a 6' b. If a ' b and c = sh(a), then c = sh(b) since closeness is transitive; thus, the order cannot be strict. Now, assume that a < b but a 6' b; fix a ' i(c) and b ' i(d). This implies that hal(a) ∩ hal(b) = Ø, but since ∗R is totally ordered, it follows that every point in one set must be less that every point in the other set. Since a < b, this gives us that i(c) < i(d), so c < d, which implies that sh(a) < sh(b). Finally, we introduce the concept of extensions. So far, we are able to relate numbers between the reals and the hyperreals. However, for the rest of the paper our focus will be on functions, so it is important to formalize the concept of transferring a function from R to ∗R. Although 3.4 deals with this matter, the introduction of these new concepts will ease notation and facilitate understanding.

Definition 4.15. Given some set A ⊂ R, the natural extension of A to ∗R is a set ∗A ⊂ ∗R defined as follows: ∗ [(rn)]U ∈ A if, and only if, {n ∈ N | rn ∈ A} ∈ U Again, we appeal to the intuition behind ultrafilters to claim this extension is ‘natural,’ by saying that a hyperreal should belong to the set if the set of entries in the sequence that belong to the original set is ‘big.’ Note also that i(A) ⊂ ∗A; however, there are also new hyperreals that belong to ∗A, a property that will actually prove to be useful. Next, we consider the natural extension of a function.

Definition 4.16. Given some function f : A → R, the natural extension of f to ∗R is a function ∗f : ∗A → ∗R defined as follows: ∗ f([(rn)]U) = [(f(rn))]U 10 Now that we have defined these notions it is important to clear up what we mean for “transfer.” We know from the transfer principle that a first order statement is satisfied by R if, and only if, it is satisfied by ∗R. However, note that we have not explicitly said how this would happen, although 3.4 makes it somewhat clear. However, in more explicit terms here is how the rule follows: we take constants in R to their image under i, we take functions in R to their natural extension and relations in R to their natural extension in ∗R. For variables, we extend the sets to which they belong and revert any of these changes when transferring back to R. We will perform transfers in the next section so the reader will be able to see some examples of these transfers then. With this last concept cleared up, we can move on to consider the concepts of non-standard analysis.

5. Continuity and Differentiability The ﬁrst concept that we will introduce is the idea of continuity. Intuitively, continuity implies that the function takes points which are ‘close’ to each other to point that are also ‘close’ to each other. The problem with this notion in R is how close is ‘close.’ A solution to this problem is the currently accepted ε − δ deﬁnition. However, in the hyperreals we have a formalized notion of closeness, which we will use to formalize the intuitive concept of continuity.4

∗ ∗ Deﬁnition 5.1. Let A ⊂ R, a function f : A → R is said to be micro-continuous at x0 ∈ A if x ' x0 implies f(x) ' f(x0) for all x ∈ A.

Note that if hal(x0) ∈ A, then f is micro-continuous at x0 if, and only if, f(hal(x0)) ⊂ hal(f(x0)). The ﬁrst important result of non-standard analysis follows:

Theorem 5.2. A function f : A → R is continuous at c ∈ R if, and only if, the natural extension of f is micro-continuous at i(c).

Proof. ⇒ Assume that f is continuous at c ∈ R and let x0 ' i(c). This implies that + + (∀ε ∈ R )(∃δ ∈ R )(∀x ∈ A)(|x − c| < δ → |f(x) − f(c)| < ε) Next, ﬁx an arbitrary ε ∈ R+, so there is a δ ∈ R+ that satisﬁes the condition above. Applying the transfer principle we get that ∗ ∗ ∗ (∀x ∈ A)(|x − i(c)|

Since x0 ' i(c), we have that |x0 − i(c)|

Theorem 5.3. (Intermediate Value Theorem) Let f :[a, b] → R be a continuous function on [a, b]. Then, for every point d between f(a) and f(b) there exists a c ∈ (a, b) such that f(c) = d. Proof. The theorem is trivial if f(a) or f(b) are equal to d, so assume without loss of generality that f(a) < d < f(b). Next consider the following statement: + 0 0 (∀ε ∈ R )(∃n ∈ N )(f(a + nε) ≤ d ∧ ((∀m ∈ N )(m > n → f(a + mε) > d)))

4The paper assumes that the reader is comfortable with the basic notions of standard single variable calculus. 11 This statement tells us that given a positive real number, there is a maximal natural number, which might be zero, such that f(a + nε) ≤ d. By transfer we get that ∗ + ∗ 0 ∗ ∗ 0 ∗ (∀ε ∈ R )(∃n ∈ N )( f(i(a) + nε) ≤U i(d) ∧ ((∀m ∈ N )(m >U n → f(i(a) + mε) >U i(d)))) i(b − a) Given this, set ε = for some ω ' ∞ and note that ε ' 0. Thus, there exists some n ∈ ∗ 0 satisfying ω N the condition above. Note that in this case n is unlimited but n < ω. Let c = sh(i(a) + nε) and note that a < c < b; we claim that f(c) = d. To show this, notice that i(a) + nε ' i(a) + (n + 1)ε since their difference is just ε, which implies that c = sh((a) + (n + 1)ε) since closeness is transitive. Thus, since ∗f is micro-continuous, we have that ∗f(i(c)) ' ∗f(i(a) + nε) ' ∗f(i(a) + (n + 1)ε), but we also have that ∗ ∗ f(i(a) + nε) ≤U i(d) < f(i(a) + (n + 1)ε) However, this also implies that ∗f(i(a) + nε) ' i(d). By transitivity we get that ∗f(i(c)) ' i(d) and, thus, ∗ f(c) = d since f is f applied pointwise. Theorem 5.4. (Extreme Value Theorem) Let f :[a, b] → R be a continuous function. Then there exist c, d ∈ [a, b] such that f(c) ≤ f(x) ≤ f(d) for all x ∈ [a, b]. Proof. We will only show that f attains a maximum, since the proof that it attains a minimum is similar. Consider the following statement: b − a b − a (∀n ∈ )(∃m ∈ ∧ m ≤ n)(∀k ∈ ) k ≤ n → f a + k ≤ f a + m N N N n n This implies that any partition of [a, b] into equal subintervals has a partition point that is maximum. By transfer we have that i(b − a) i(b − a) (∀n ∈ ∗ )(∃m ∈ ∗ ∧ m ≤ n)(∀k ∈ ∗ ) k ≤ n → ∗f i(a) + k ≤ ∗f i(a) + m N N U N U n U n Fix ω ∈ ∗N, ω ' ∞; then there exists some m ∈ ∗N satisfying the conditions above. Define d = b−a sh a + m n , we claim that this is the desired point. Notice that since f is continuous we have that ∗ b−a ∗ f a + m n ' f(i(d)). Now, we must show that f(d) is actually maximal; fix some x ∈ [a, b] and notice that b − a b − a (∃n ∈ ∗ 0) a + n ≤ i(x) ∧ (∀m ∈ ∗ 0) n < m ≤ ω → i(x) < a + m ≤ i(b) 0 N 0 ω U N 0 U U U ω U This statement follows from transfer. Note that it only states that x exists between two points of the ∗ b−a partition, which is clearly true in R and it transfers into R. Note that this implies that i(x) ' a + n0 ω , ∗ ∗ b−a which also gives us that f(i(x)) ' f a + n0 ω since f is continuous. However, note that n0 ≤U ω, so b − a b − a ∗f(i(x)) ' ∗f a + n ≤ ∗f a + m ' ∗f(i(d)) 0 ω U n This implies that f(x) ≤ f(d) since both x and d are real. Since this holds for all x ∈ [a, b], the proof is complete. Next, we will give a non-standard definition of the derivative and prove some properties using non-standard analysis. Intuitively, the derivative is supposed to be the slope of the tangent line at some point x0. However, the problem is that to define such a line we require at least two points, while we only have one. In standard analysis, this is resolved by letting a second point x approach the first point using limits and looking at the slope of the line when the two points are arbitrarily close together. In non-standard analysis, we use a similar approach but instead make use of infinitesimals to define the second point.

Definition 5.5. Let f : A → R be a function. We say that f is differentiable at x0 ∈ A if there exists a unique L ∈ ∗R such that for every nonzero infinitesimal ε we have ∗f(i(x ) + ε) − ∗f(i(x )) 0 0 ' L ε 0 If so, we define the derivative of f at x0 to be the shadow of L, meaning f (x0) = sh(L). 12 While the definition should appear intuitive to the reader, we still must show that it is equivalent to the standard definition using limits.

Theorem 5.6. A function f is differentiable at x0 as defined in 5.5 if, and only if, it is differentiable at x0 in the usual sense. Furthermore, if it is differentiable, the derivatives are the same.

0 ∗ Proof. ⇒ Assume f is differentiable at x0 ∈ A as defined in 5.5 such that f (x0) = sh(L) for some L ∈ R. Fix some ε ∈ R and consider some δ ∈ ∗R+ with δ ' 0, so if i(0) ≤ |µ| < δ, then µ ' 0. Since f is differentiable we get that ∗ ∗ f(i(x0) + µ) − f(i(x0)) − sh(L)

As in standard analysis, we can redeﬁne the derivative by looking at points inﬁnitely close to x0.

Definition 5.7. Let f : A → R be a function. We say that f is differentiable at x0 ∈ A if there exists a ∗ unique L ∈ R such that for every y ∈ hal(x0) \{x0} ∗f(y) − ∗f(i(x )) 0 ' L y − i(x0) 0 If so f (x0) = sh(L). The fact that these two definitions are in fact equivalent should be evident, since the fact that y ∈ hal(x0) \{x0} implies that i(x) − y is a nonzero infinitesimal. With this new definition of the derivative we can easily motivate Leibniz’s notation for derivatives, albeit with some abuse of notation. Let dx be a nonzero infinitesimal, then we can define ∆f as the increment of f at some point x by setting ∆f = ∗f(i(x) + dx) − ∗f(i(x)). Notice that it follows that ∆f ' f 0(x) for all nonzero infinitesimals dx. dx 0 ∆f Furthermore, notice that since f (x) is limited and ∆f = dx dx, we have that ∆f ' 0 since the product of a limited number and an infinitesimal is infinitesimal. Note that this proves an important fact.

Theorem 5.8. If f is differentiable at x0, then f is continuous at x0. 13 And by setting df = f 0(x)dx, we see that ∆f df ' = f 0(x) dx dx Finally, of note is the fact that this definition of the derivative allows us to give proofs of all the known differentiation rules without appealing to the limits definition. However, the proofs are left to the reader.

6. Integration When integration is first introduced in a standard calculus course the intuition behind it generally proceeds as follows: we consider a bounded function f over a certain interval [a, b]; then we see that we can approximate the value of the area under the curve by means of rectangles normally by partitioning [a, b] into smaller intervals of the same length and then using the image under f of the first point or last point of this subintervals as the length of the rectangle. We quickly notice that as we make the subintervals smaller, the better our approximation of the total area is. Thus, when the subintervals have infinitesimal length, we should obtain the actual value of the area under the curve. However, there are some issues with this notion, which we will regard as the “partition problem.” To introduce this issue we will first present the common notions of integration in the standard sense. From here on out, we assume f to be a bounded function defined on some interval [a, b] ⊂ R.

Deﬁnition 6.1. A Partition of an interval [a, b] is a ﬁnite set P = {x0, x1, . . . xn} ⊂ [a, b] such that x0 = a, xn = b and xi−1 < xi for all 0 ≤ i ≤ n.

Let Mi and mi be the least upper bound and greatest lower bound of f in the interval [xi−1, xi] then we define n n X X L(f, P ) = mi(xi − xi−1) U(f, P ) = Mi(xi − xi−1) i=0 i=0 L(f, P ) is known as the lower sum and U(f, P ) is known as the upper sum. Note that L(f, P ) ≤ U(f, P ) for all partitions P . Graphically, L(f, P ) takes the smallest value of f in each subinterval as the length of the rectangle while U(f, P ) takes the highest value. Note that for any bounded function we have L(f, P ) ≤ U(f, Q) for any two partitions P and Q of [a, b] We can use these values to define integration as follows: Definition 6.2. We say that a function is Riemann Integrable on an interval [a, b] if given any ε > 0, there exits a partition P such that L(f, P ) − U(f, P ) < ε. This is sometimes know as the Cauchy Criterion for integration. Intuitively, it means that we can make the subintervals so small that the difference between the lower and upper sums is less than any value we desire. Note that this definition resembles the limit definition. This also means that over all partitions of [a, b], L(f, P ) is bounded above and U(f, P ) is bounded below; furthermore, it implies that these bounds are equal and we call this value the integral of f from a to b. The proof of this fact should be familiar to the reader.

Now, if we want to transfer this notion to nonstandard analysis, we might be tempted to make it so that the difference between any two points in the partition is infinitesimal, maybe by dividing b−a by un unlimited number ω. This would imply that P would not only be infinite but also uncountable, which brings up the question of how do you sum over uncountably many elements? This is precisely the partition problem, that it is difficult to define a rigorous way of summing over a partition which has as many elements as R itself. One solution, presented by Goldblatt in [6], is to consider only partitions where all consecutive points are equidistant (meaning xi − xi−1 = ∆x) except for the last two points whose difference might be less that ∆x. Then, we can think of L(f, P ) and U(f, P ) as functions defined on R+ determined by this particular value ∆x and extend this functions to ∗R in the natural way. However, this only masks the problem; it doesn’t solve it since it doesn’t really tell us how these values are computed or what their meaning is. Thus, we will take a slight detour, which is less intuitive but more rigorous, by introducing the notions of hyperfinite sets and hyperfinite sums. 14 ∗ Definition 6.3. Given a sequence (An) of subsets of R, we define a set [(An)]U ⊂ R by the following rule [(rn)]U ∈ [(An)]U ⇔ {n ∈ N | rn ∈ An} ∈ U A set A ⊂ ∗R is called internal if it can be defined in this way. Otherwise, we say that A is external. ∗ Furthermore, note that if An = A ⊂ R for all n ∈ N, then A = [(An)]U. ∗ Definition 6.4. An internal set [(An)]U ⊂ R is called hyperfinite if {n ∈ N | An is finite} ∈ U. In this ∗ case, we define the cardinality of [(An)]U as follows: |[(An)]U| = [(|An|)]U ∈ N. ∗ Definition 6.5. Given an internal set A = [(An)]U ⊂ R and a function f such that An is in the domain of f for all n ∈ N, we can define the hyperfinite sum of f on A as follows " !# X ∗ X ∗ X f(x) = f([(xn)]U) = f(xn)

x∈A [(xn)]U∈A xn∈An U

Note that this definition makes sense since An is finite for most n ∈ N. Also note that this sum respects most of the properties normal sums do in R. This definition also follows the same idea we’ve been using all throughout nonstandard analysis of defining concepts in ∗R as the pointwise application of its analogoes in R. With this notion, we can begin to define non-standard integration.

Definition 6.6. Given a positive infinitesimal dx = [(εn)]U, for each εn we define a partition Pn by noting that + (∀ε ∈ R )(∃k ∈ N)(a + kε ≤ b ∧ ((∀m ∈ N)(m > k → a + mε > b))) so Pn = {a, a + εn, a + 2εn, . . . , a + knεn, b}, where kn is the corresponding k, as defined above, for εn. This gives us a hyperfinite partition P = [(Pn)]U of [a, b]

Note that if εn > a − b, then Pn = {a, b}; similarly, if εn < 0, we also define Pn = {a, b}. Furthermore, for all 0 ≤ i ≤ kn − 1 we have that xi − xi−1 = εn and xn − xn−1 ≤ εn. This will allow us to sum in the usual sense. First let’s redefine L(f, P ) and U(f, P ) as functions of ε ∈ R+ by saying that L(f, ε) = L(f, P ) where P is the partition defined in 6.6, same for U(f, ε). Therefore, we can extend this functions as follows:

Deﬁnition 6.7. Given a positive hyperreal dx = [(εn)] we deﬁne the extensions of the upper and lower sums of f on [a, b] as " kn !# ∗ X U(f, dx) = Min εn = [(U(f, εn))]U in=0 U

" kn !# ∗ X L(f, dx) = min εn = [(L(f, εn))]U in=0 U ω X b − a Note that we would like ∗U(f, dx) to be defined as M dx where dx = . This definition captures i ω i=0 that sense by letting An, as defined in 6.5, be the set containing all the Mi for the partition Pn. Now, we can define integrability: Definition 6.8. We say that a function f is ∗Riemann integrable on [a, b] if given any positive infinitesimals dx, we have that ∗U(f, dx) ' ∗L(f, dx). Furthermore, we say that Z b f = sh (∗U(f, dx)) = sh (∗L(f, dx)) for any nonzero infinitesimal dx. a It might not be evident to the reader why this should be the value of the integral, or at least it does not seem to follow intuitively. First of all, note that since dx ' 0, it follows that the sequence (εn) converges to 0; this implies that given any positive real number δ, there are infinitely many elements of the sequence which are less than δ. In that sense, we are making the partitions infinitesimals. Nonetheless, a proof might be more convincing to the reader. But first, it is convenient to redefine the Cauchy Criterion in terms of positive real numbers. 15 Definition 6.9. We say that a function is Riemann integrable on an interval [a, b] if + + + (∀ε ∈ R )(∃δ ∈ R )(∀∆x ∈ R )(∆x < δ → U(f, ∆x) − L(f, ∆x) < ε) The proof that this is equivalent to the Cauchy Criterion should be easy for the reader to follow, so we will only give an overview. If 6.9 is satisfied then the Cauchy Criterion is satisfied by creating a partition using 6.6. Conversely, if the Cauchy Criterion is satisfied, then let δ = min(xi − xi−1) over all possible i in the partition P that the criterion guarantees. It is left to the reader to show that this δ is sufficient for 6.9. Theorem 6.10. A function f satisfies the Cauchy Criterion for integration if, and only if, f satisfies 6.8. Proof. ⇒ Fix ε ∈ R+ then there is a δ ∈ R+ such that + (∀∆x ∈ R )(∆x < δ → U(f, ∆x) − L(f, ∆x) < ε) . By transfer we get that ∗ + ∗ ∗ (∀dx ∈ R )(dx

R b + + Proof. ⇒ Let I = a f by means of 6.11. Fix ε ∈ R ; thus there is a δ ∈ R such that by transfer we get that ∗ + ∗ ∗ (∀dx ∈ R )(dx

R b ∗ ∗ + + ∗ + ⇐ Let a f = sh( U(f, dx)) for some positive inﬁnitesimal dx ∈ R . Fix ε ∈ R and let δ ∈ R be a positive inﬁnitesimal. This implies that ∗ + ∗ ∗ ∗ ∗ (∀dy ∈ R )(dy

7. Conclusion Non-standard analysis provides a new take on the usual concepts of analysis. While we only explored the basics of single variable calculus, most of analysis can be developed in this matter; Robinson and Goldblatt provide a more thorough study of non-standard analysis, including more advanced concepts like measure theory and the Lebesgue integral, and are a must read if the reader wants to become more familiar with these concepts. Acknowledgments. I would like to thank my mentor Jacob Perlman for his help and advice in the construction of this paper. I would also like to thank Professor Peter May for making the 2015 Math REU possible and all the lecturers, mentors and students involved for making the REU such an enriching experience.

References [1] William Weiss and Cherie D’Mello, Fundamentals of Model Theory, University of Toronto, 1997, http://www.math.toronto.edu/weiss/model_theory.pdf. [2] William DeMeo, First-Order Logic (Predicate Calculus), https://williamdemeo.files.wordpress.com/2013/01/cs-firstorder.pdf. [3] Alex Kruckman, Notes on Ultraﬁlters, Toolbox Seminar 11/7/12, https://math.berkeley.edu/ kruckman/ultrafilters.pdf. [4] Klaus Kaiser, Ultraproducts, University of Houston, http://www.math.uh.edu/ klaus/ultraproducts.pdf. [5] Sabine Koppelberg, Logik und Mengenlehre (Chapter 2), Freien Universitt Berlin, http://www.mi.fu-berlin.de/math/groups/ag-logik/Lehre/UST-chapter02.pdf. [6] Robert Goldblatt, Lectures on the Hyperreals, An introduction to Nonstandard Analysis, Springer, 1998. [7] Robert A. Herrman, Nonstandard Analysis - A Simpliﬁed Approach -, http://arxiv.org/pdf/math/0310351.pdf, 2003. [8] Abraham Robinson, Non-standard Analysis, Princeton University Press, 1974.